1

Efficient Thinking

Can a fixed AI model become more capable without becoming larger?

❄
Frozen model
(same weights)
→
Read-only head
(experience)
→
Better search
decisions

Eight measured papers, one frozen model at a time. Every number on these slides is a measured result from the published papers; nothing is projected.

→ / ← or scroll · one slide per step of the argument

The series in one line

Same question from every angle: how much useful capability can a fixed intelligence get per unit of computation, and what moves that number?
2
I – IISearch
→
III – VEvaluation & feedback
→
VI – VIILabels & the internal ceiling
→
VIIIExperience
→
IXSymmetry
→
XRepresentation
Hold the model fixed. Not one parameter changes in any paper; capability is scored by an external verifier and charged per unit of computation.
Change one thing. Search, the evaluator, the judge, labels, the internal ceiling, experience, symmetry, representation: one variable per paper.
Where this sits. Efficient reasoning asks the same question from the compute side (fewer tokens). This series answers from the experience side, weights frozen. Two methods on one axis.

I → II → III: search, its ceiling, and the price of judging

Papers I–III establish the frontier and its limits: search moves a fixed model along its quality–compute frontier, and only external information raises the ceiling.
3
I · Measuring What Capability Costsclosed · archived (DOI)
Does where you spend computation change what a fixed model can do?
Yes. A 14 MB chess net with its evaluator fixed gains about 650 Elo from search alone: capability from allocation, not size.
Fixed evaluator
→
Search
→
+650Elo
Key results
  • Evaluator fixed; search depth the only variable
  • +286 Elo at 800 playouts on the same ladder; ~650 at the full budget
  • Internal ladder Elo, ±~100 systematic near the top rung
II · Where Search Pays and Where It Can'tclosed · archived (DOI)
Does sampling a fixed model more and more keep improving the answer?
No. Sampling saturates at a consensus of the model's own errors; only an external verifier lifts the ceiling.
❄
Frozen model
→
N samples
→
✓
Verifier breaks
the ceiling
Key results
  • Consensus converges on correlated errors
  • A perfect verifier over the same samples keeps rising
  • VIII's experience is labelled by such a verifier: "verified history"
III · Efficient Judgingclosed · archived (DOI)
A verifier is not free. When is a model-based judge worth its compute?
Only in one corner. An LLM judge beats free majority vote only for weak policies and small candidate fields, once its cost is charged.
✓
LLM judge
vs
Majority vote
→
judgingper unit compute
is the capability
Key results
  • Eight cells survive significance: a strong judge (14B–72B) selecting for a weak policy (0.5B–1.5B)
  • Charged honestly, the judge often loses to the vote
  • VIII's compute accounting carries the verifier's cost because III showed it must
1. Allocation is capability (I)search substitutes for size
2. The ceiling is the evaluator (II)only external truth lifts it
3. Judging has a price (III)capability is judging per unit of compute

VII: The Elicitation Gap — nothing is free inside

How much does a model know that it does not say? Under a pre-registered forced-choice definition, the claim fails at every size.
4
VII · The Elicitation Gapdraft v0.5 · E-E complete at four sizes
Is there a large reserve of knowledge in the state that the model cannot express?
No, not in this setup. Made to answer, the model says what its state knows. The 48-point "gap" of the first draft was ties scored as ignorance; it is withdrawn.
❄
Frozen model
→
Probe on the
hidden state
vs
✓
Judge made
to answer
Key results
  • Forced judge rises 0.53 → 0.73 from 1.5B to 32B
  • Probe at matched depth stays flat: 0.68 · 0.73 · 0.70 · 0.72
  • Tie rate collapses 99% → 21%: the abstention was format, and scale fixed the format
  • Decodable ≠ usable ≠ controllable — VIII-g met the same wall from the other side
Scale buys expression, not representationmeasured, four sizes
0.40.50.60.60.70.81.5B7B14B32Bforced judgeprobe, matched depth
Why it is the hinge of the series
  • If nothing is free inside a fixed model, any gain at fixed intelligence must come from outside the weights
  • The state can be read directly — and Paper VIII reads it
  • A negative that corrected its own headline, dated and kept

VIII: From Use → Accumulation → Acquisition

Same intelligence (frozen weights), same model, same read-only head — different experience. What matters and why.
5
VIII-a · Experience Priorsv1.0.2 · closed · reproduced
Can verified historical experience improve the search of a frozen intelligence?
Yes — a frozen model learns where to look from its own verified past without changing its weights, moving its quality–compute frontier.
Verified history
(successful episodes)
→
Experience controller
(linear head, read-only)
→
Better search
decisions
081624324022.3%basegiven the head's compute31.7%frozen headon verified history
Key results
  • Matched compute: 31.7% vs 22.3% (+9.3 [+3.7, +15.0], p 0.002)
  • Equal quality (24.0% both) at 2.32× less compute
  • Six alternative carriers of experience fail the constraint
  • Second search structure (SQL repair), head re-fitted: +12.3 [+6.7, +18.0]
  • Reproduced on a second machine: +8.0 [+3.0, +13.3]
VIII-b · Accumulationdraft v0.3 · every arm run
Does experience generated under generation 1 produce an even better generation 2?
No clear accumulation — self-generated experience improves somewhat but does not separably compound against the loop's own noise.
E₀ (base)
→
❄
Frozen model
+ E₁ head
→
E₁ (self-experience)
081624324022%no experience(base)26.8%self-experience(generation 1)
Key results
  • Base 22.0% · self-experience 26.8% (+4.8, not separable)
  • Does not clear the registered significance and stability bars
  • Duplication revealed the saturation; disjoint problems showed it was not the cause
  • The limit is the region of state space the policy keeps returning to
VIII-c · Where the Experience Is Acquiredv1.0.12 · closed · reproduced
Does the source and method of acquisition change the value of experience?
Yes — how experience is acquired changes its value sharply, breaking the self-experience plateau without touching the model's intelligence.
Self-experience
(own states)
Other-agent
(where a second
agent stood)
Forced-look
(one look before
the first guess)
0102030405022%noexperience26.8%self-experience35.4%other-agentstates41.2%forced-lookexperience
Key results
  • No experience 22.0% · self 26.8% · other-agent 35.4% · forced-look 41.2%
  • 132 forced-look groups (32.6%) indistinguishable from 290 other-agent groups (35.4%)
  • The cheap proxy tracks amount of experience, not origin
  • Reproduced from scratch on a second machine within the pre-registered bar
  • Still open: why it works (coverage vs. representation)
1. Experience Use (VIII-a)verified experience shifts a frozen intelligence's frontier
2. No Automatic Accumulation (VIII-b)more self-generated experience does not clearly compound
3. Acquisition Method Matters (VIII-c)where and how experience is acquired changes its value
Three separable variables for efficient improvement  ·  Intelligence (fixed weights)  |  Experience acquisition (where the agent goes)  |  Experience use (how it guides search)

VIII-d · e · f: select it? retrieve it? does it transfer?

The follow-up arms ask whether the value lies in choosing experience, in showing it to the model, or in the method itself.
6
VIII-d · Selecting Experiencemeasured 2026-10-01/02 · five arms
Does choosing which past states the head is fitted on beat a random draw from the same pool?
No. Coverage, readability and a mix-matched control all sit within the design's ~5-point resolution of each other and of a random draw.
The pool
→
rule?
→
Head on 132 states
Key results
  • No selection rule beats a random draw from the same pool
  • The +3 to +8 that looked like selection was the pool
  • Registered before each draw
VIII-e · Retrieval Against the Headmeasured 2026-10-01 · fifteen games
Does putting past experience into the prompt do what the read-only head does?
No. Prompt retrieval (k=1, matched bank) adds nothing and costs 5.9% more input; the head beats it in every game.
Retrieved into
the prompt
vs
Read by
the head
04812162013 ptshead over retrievalsmallest game15 ptshead over retrievallargest game
Key results
  • Retrieval against base is a null
  • Head +13 to +15 points over retrieval, every game
  • A 23-point offline advantage of one label converted to no in-loop advantage
VIII-f · A Third Familymeasured 2026-10-03 · registered before compute
Is the head's use of experience a property of one model, or of the method?
Of the method. Yi-1.5-9B-Chat at 4-bit, under the same protocol, with its own head from its own history.
❄
Family 1
❄
Family 2
❄
Family 3
05.611.216.822.4289.3 ptsfamily 1matched compute14.3 ptsfamily 2disjoint 30021.3 ptsfamily 34-bit, game 1
Key results
  • Head over base +21.3 / +22.0 / +23.0, three of three
  • Head over retrieval +20.7 / +22.0 / +22.3
  • Retrieval against base: a null. The verdict transfers; the size does not
Not selectionthe pool, not the rule
Not the promptthe state, read directly
Not one modelthree families, one verdict

VIII-g · h · i: what the head can be, which label makes it, and at 4-bit

Three controls on the mechanism: steering, the label, and precision.
7
VIII-g · Steering From the Verifier's Labelmeasured negative 2026-10-03
Can the verifier's label steer the model's state directly, as the adaptive-steering literature does?
No, at its first gate. The separability precondition is not met for this label at layers 18 and 27.
Verifier label
→
d′ 0.74 / 0.77against a bar of 2
Key results
  • Less separable than the agent's own pick
  • Does not conflict with VIII-f: within-task d′ excludes the cross-task structure a head uses
  • Decodable information is not steerable information
VIII-h · Which Label Makes the Better Headmeasured 2026-10-05 · criterion ruled symmetric before any count
Does the verifier's label beat the agent's own pick as the thing a head is fitted on?
Yes. On nine disjoint task sets with a pooled estimator, the verifier-label head wins nine of nine.
✓
Verifier label
vs
Agent's own pick
04.89.614.419.22416.6 ptsverifier headover own-pick head
Key results
  • +16.6 points, interval [+14.6, +18.5]
  • Nine of nine games positive
  • Useful experience carries information the agent did not hold in its own behaviour
VIII-i · The Quantisation Controlmeasured 2026-10-05 · second read signed
Does the family-1 result survive the same 4-bit precision the third family ran at?
Yes. "Survives 4-bit" is now a measurement on both families.
❄
Family 1, 4-bit
→
Same head,
same protocol
07.214.421.628.83626.9 ptshead over base28.8 ptshead over retrieval
Key results
  • Both criteria, three of three
  • Retrieval against base: a null for the third time
  • The size of the gain differs by family; the verdict does not
Cannot steerthe label is not separable enough
The verifier's label wins+16.6, nine of nine
Survives 4-biton both families

The claim, the method, and where in the chain it works

What the series says, in the sentences the papers use.
8
The claim. When intelligence is held fixed, the value of experience depends strongly on how that experience was acquired, not merely on how much experience is stored or how it is retrieved. Every clause is a measured result: acquisition (VIII-c), amount (VIII-b), retrieval (VIII-d, VIII-e).
The method. The model's parameters are its thinking and are never changed; everything else that can move search — what the model is given to read, where that experience is collected, how it is applied at a decision — is varied one at a time and measured. Holding the thinking fixed is what lets each gain be attributed to experience.
Where it works. The head decides where the next inspection goes, which is where the next unit of computation goes: a learned search prior, closer to a tree search's selection rule than to an action classifier. Its measured effect is on allocation: equal quality at 2.32× less compute.
acquisition policy→experience→search policy→computation→outcomewith the intelligence fixed at every link

What the method risks, and how that is measured

Changing parameters carries a cost nobody can quantify. This series never changes one, so that cost is zero by construction. Its own risk is stated and measured.
9

The risk. The head chooses among candidates the model itself proposed, so it can steer search away from a place where the vanilla model would have succeeded.

The measurement. Problem by problem, on the same problems: the base-only cell of a paired four-cell table (both solve · head only · base only · neither), published beside every headline, with both denominators.

Across the closed arms: 1.2% to 12.2% of the task set; 11.7% to 59.8% of the problems the vanilla model itself solved. The regression does not track the gain: the largest gains carry the smallest regressions. A head is applied only within the family it was fitted on; outside it the vanilla model runs unchanged.

The regression cell, two denominatorsderived 2026-10-05 from the episode files
014284256701.2%smallestof task set12.2%largestof task set11.7%smallestof base's own59.8%largestof base's own
  • Every published mean reproduced from the same files
  • The second denominator is the one that prices the behaviour
  • A head outside its family: registered as the next control

Still ahead — registered designs, not findings

Every paper opens with a state line (measured, registered, concept, idea) so a concept is never read as a finding. Withdrawn claims stay in the record, dated.
10
IV — Search where taste is the evaluator
Poetry and music, where no checker can score the answer. Machine arms measured; rater arms running.
V — The exchange rate of feedback
Improvement per bit of human correction. Concept.
VI — The label ceiling
Self-play stops where its labels stop being true. Measured, unwritten.
IX — Symmetry
Solve one case, cover its whole orbit. Concept.
X — Machine-native representation
Is language the efficient form for a machine to reason in? Concept with a plan.
Registered next controls
A head applied outside its family; the acquisition rule on a second task family.

Reproduce it, or break it

Every table carries its command and interval; result artefacts name their instruments by hash; readings are written before runs; locks are second-read before scores.
11
git clone https://github.com/LouayAlsakka/efficient-thinking && cd efficient-thinking
python experience/g13b_score.py --dir experience/results/viiic --plan viiic7 --out /tmp/viiic.json   # our bytes, minutes
bash experience/reproduce_8c.sh                                                                      # your bytes, ~4.5 h
bash experience/reproduce_matched_compute.sh                                                         # base at the head's compute vs the head
Review is public and runs through the repository's issues; the author will run any experiment you name. The four places the author thinks each paper is weakest are on the challenge page.
To cite: Papers I–III doi 10.5281/zenodo.21520992 · Papers VIII and VIII-c doi 10.5281/zenodo.23005271 · Author: Louay Alsakka
Three separable variables for efficient improvement  ·  Intelligence (fixed weights)  |  Experience acquisition (where the agent goes)  |  Experience use (how it guides search)

Site build after e567b61 · 2026-10-07 13:27Z